New AMD Radeon graphics card PCB photos with GDDR6 memory - NAVI Spotted?

Discussion in 'Frontpage news' started by Hilbert Hagedoorn, Apr 27, 2019.

  1. Hilbert Hagedoorn

    Hilbert Hagedoorn Don Vito Corleone Staff Member

    Messages:
    48,392
    Likes Received:
    18,565
    GPU:
    AMD | NVIDIA
  2. Clawedge

    Clawedge Guest

    Messages:
    2,599
    Likes Received:
    928
    GPU:
    Radeon 570
    [​IMG]

    I know its just the PCB, but the pressure has been building up for so long, it just came.....out.
     
  3. HybOj

    HybOj Master Guru

    Messages:
    394
    Likes Received:
    325
    GPU:
    Gygabite RTX3080
    Dude learn some manners :D
    One would say... that escalated quickly .)
     
    Keitosha, Rich_Guy and Clawedge like this.
  4. airbud7

    airbud7 Guest

    Messages:
    7,833
    Likes Received:
    4,797
    GPU:
    pny gtx 1060 xlr8
    Clawedge likes this.

  5. Fox2232

    Fox2232 Guest

    Messages:
    11,808
    Likes Received:
    3,371
    GPU:
    6900XT+AW@240Hz
    Here's the thing:
    Fury X has double memory bandwidth of RX-580, but is only around 20% faster.
    Radeon 7 has double memory bandwidth of Vega 64 and is again only around 20% faster.

    Now, 256bit GDDR6... depending on clock it will be up to 512GB/s theoretical maximum bandwidth, which would be double of RX-580 and theoretically enable double performance... if GPU is strong enough to actually use this bandwidth. More realistically memory clock will be lower and bandwidth will be around 420~460GB/s as AMD is probably not aiming for most expensive memory chips.
    - - - -
    Is this Navi card aiming at 1.7~2x performance of RX-580? Maybe, but there is quite some chance that it will be only 1.5times as fast. And AMD simply designed both GPU and PCB from experience with Polaris where final cards have good 40% higher clock, but are held back by memory bandwidth limitation.
    And in best case scenario this card is going to have better IMC, caches in "CU"s and improved memory command querying... Enabling 10~20% higher achievable GPU performance at same bandwidth as Polaris.
    - - - -
    My guess from this is that:
    - Worst case scenario: Vega 56 performance (Unlikely as it would require regression in technology or small GPU with only 6~7B transistors... too cheap GPU with expensive memory.)
    - Expected performance: RTX 2070 +-5% (Very likely)
    - Optimistic situation: RTX 2080 +-5% (Unlikely)
    - Great Design choices but higher price: RTX 2080 +10~15% (Highly unlikely)
    - - - -
    This PCB has basically RTX 2070/2080 memory bandwidth potential. But I do not exactly see AMD planning for GPU with more than 12 Billion transistors as they likely want those cards to have price impact.
    It may be possibly smaller GPU like 8.5~9B transistors and utilize higher clock (~1.7GHz), Which would be matching performance improvement from RX-580 to RTX 2070.
     
    Maddness likes this.
  6. RooiKreef

    RooiKreef Guest

    Messages:
    410
    Likes Received:
    51
    GPU:
    MSI RTX3080 Ventus
    Lmao! Best post I have ever seen on the Guru3d forums!
     
    Keitosha, Silva and Clawedge like this.
  7. kings

    kings Member Guru

    Messages:
    158
    Likes Received:
    131
    GPU:
    GTX 980Ti / RX 580
    I wanted to edit, but I accidentally deleted my previous post. :confused:

    Regarding the RTX line, it depends, RTX 2060 for example has similar performance to the Vega 64 and also have similar prices.

    It's not Nvidia's fault that with heavily cut down chips, it can perform equal or better than AMD cards. At most, that says a lot about AMD's inability in recent years.

    Nvidia being the dominant brand on the market, has no interest at this point in being disruptive in pricing.
     
    Last edited: Apr 28, 2019
  8. Fox2232

    Fox2232 Guest

    Messages:
    11,808
    Likes Received:
    3,371
    GPU:
    6900XT+AW@240Hz
    It is just name. And therefore not exactly objective. Yes nVidia seemingly assigned lower tier chip to higher tier card and asked higher price, but here is some other way to look at it...
    Number of transistors card in given tier has per $ (MSRP):
    Old.Gen => New.Gen : Old.TransPer$[Millions] => New.TransPer$[Millions]
    1060 => 2060 : 14.72 => 30.95
    1070 => 2070 : 19 => 21.64
    1080 => 2080 : 12 => 19.46
    1080Ti => 2080Ti : 16.88 => 18.62
    ... There are apparently cut down chips which skews down perspective a bit, but in each performance tier nVidia delivered bit more transistors per $ they ask for final card. Yes, they are mostly invested into new features from which some are still vaporware, some proven to be worse than expected, and some are borderline marketing tool.

    But here is normalization of Full chips as both lineups have quite some cut down variants filling ranks:
    1060 => 2070 : 14.72 => 21.64
    1080 => 2080 : 12 => 19.46
    1080Ti => 2080Ti : 16.88 => 18.62

    And here are full chips of similar transistor count... normalized to same transistor count and multiplier shows how many more transistors new gen delivers per $:
    1060 => 1650Ti : 1.56x (If 1650Ti has $219 MSRP)
    1080 => 1660Ti : 1.81x
    1080Ti => 2070 : 1.17x (Yes, 1080Ti transistor count is closet to 2070 than to 2080.)
    Yes, people here do not want 1660Ti instead of 1080 as later is 20% faster... where did all those TMU/ROPs go on same sized GPU? Into FP16 capability like in Vega.

    Normalizing numbers above even to performance:
    1080 => 1660Ti : 1.64x
    1080Ti => 2070 : 1.19x

    I would say that from nVidia's technological and economical point of view, they delivered more and cheaper.
    But from our side, we really do not care about nVidia wasting transistors as AMD did. Clients want to have upgrade or at least side grade which would be quite cheaper. Clients want higher performance per $ without having to downgrade.
    And that did not happen. In this metric, stock 2070 has mere 2% higher performance per transistor than 1080Ti. 2070 has lower TDP and OCed delivers more, but there are 2 higher Full GPUs above it which have 215W and 250W TDP. (And invading their power draw range out of box would make 2070 unattractive from this perspective.)
    And 1660Ti would be downright downgrade in terms of performance per transistor as it is 16% slower than 1080, but has only 8% fewer transistors. (That's around 10% higher performance per transistor on 1080 side. But 1080 gain ate like 50% more power... not Turing improvement, but manufacturing technology.)
    = = = =
    TL;DR: Turing is improvement on technological level over Pascal. Those cards do deliver better performance per watt. Better performance per $. They are not as attractive due to not very good performance per transistor which kind of remained same while nVidia added 2 bigger GPUs above 1080Ti.
    This resulted in bigger difference in target TDPs over lineup against older cards with same transistor count. Which is on one hand good, but gets to show that we really do not care that much about power draw outside of "AMD vs. nVidia" arguments.
     
    Aura89 likes this.
  9. mockingbird2

    mockingbird2 Guest

    Messages:
    1
    Likes Received:
    0
    GPU:
    GeForce GTX 1080
    If you need any more nails in the coffin of AdoredTV's fake "leaks", here they are.

    If the supposed "Radeon RX 3080" has a TDP of 150W, why would it have two 8 pins connectors and have VRMs on par with the reference Vega?

    It wouldn't.
     
  10. Yakk

    Yakk Guest

    Messages:
    144
    Likes Received:
    21
    GPU:
    Graphics card
    Looks like an engineering/debugging board to me, it should have a lot of extra unneeded stuff for production.
     
    Fox2232 likes this.

  11. Fox2232

    Fox2232 Guest

    Messages:
    11,808
    Likes Received:
    3,371
    GPU:
    6900XT+AW@240Hz
    Or it can be really just reference board for top 3 models. Top model may be like 250W.
    Then weaker model 205W, and weakest model on this PCB may be 160W.
    Those may be based on same chip, just cut down for lower variants. And in similar way memory may be 6/7GB on lower models. In similar fashion as nVidia has unused memory places on PCBs. And some VRMs may remain unused on lower models too.

    There are many possibilities. And with all the unknowns, we do not really know what is AMD cooking. There seems to be somehow unsuccessful tapeout at end of last year. But by now they had enough time to get things in order and do 2 more. We do not even know if it is still on older 7nm or on newer one.

    I personally like to speculate, but only for fun of thinking over all the variables which are unknown and where they may end. But for myself, I know that even information provided by AMD themselves may be outdated. And plans may have changed as now we are much closer to PS-next release with something like 10~14 TFLOPs GPU. Now with date set to 7th of July, we may get to see many more GPUs than AMD could have introduced in January.
     
  12. sykozis

    sykozis Ancient Guru

    Messages:
    22,492
    Likes Received:
    1,537
    GPU:
    Asus RX6700XT
    I'm tired of leaks, rumors and speculation.....give me the card already...
     
    airbud7 likes this.
  13. airbud7

    airbud7 Guest

    Messages:
    7,833
    Likes Received:
    4,797
    GPU:
    pny gtx 1060 xlr8
    you just bought a new card!...how many you need?....:p
     
  14. mockingbird1

    mockingbird1 Guest

    Messages:
    7
    Likes Received:
    0
    GPU:
    8 GB
    What are you taking about?

    There is no evidence that there was supposed to be tape-out at the end of last year.
     
  15. Fox2232

    Fox2232 Guest

    Messages:
    11,808
    Likes Received:
    3,371
    GPU:
    6900XT+AW@240Hz
    Read carefully.
     

  16. fantaskarsef

    fantaskarsef Ancient Guru

    Messages:
    15,696
    Likes Received:
    9,574
    GPU:
    4090@H2O
    Cool. Waiting for benchmarks anyway.
     
  17. Supertribble

    Supertribble Master Guru

    Messages:
    978
    Likes Received:
    173
    GPU:
    Noctua 3070/3080 FE
    Radeon VII is expensive as you say but I wouldn't really call it a rip-off as it's expensive to manufacture. I doubt AMD is making much money from it.
     
  18. Undying

    Undying Ancient Guru

    Messages:
    25,343
    Likes Received:
    12,754
    GPU:
    XFX RX6800XT 16GB
    Looking at the board RTX2070/2080 performance. Price it at 350$ and it will disrupt the market.
     
    Last edited: Apr 29, 2019
  19. vbetts

    vbetts Don Vincenzo Staff Member

    Messages:
    15,140
    Likes Received:
    1,743
    GPU:
    GTX 1080 Ti
    So if it indeed is GDDR6, that means in Navi they were able to lower the power consumption enough to throw on GDDR6 instead of HBM. That's pretty awesome to look forward to.
     
  20. Rich_Guy

    Rich_Guy Ancient Guru

    Messages:
    13,138
    Likes Received:
    1,091
    GPU:
    MSI 2070S X-Trio
    :D
     

Share This Page